[FFmpeg-devel] [PATCH V2 3/3] dnn: convert tf.pad to native model in python script, and load/execute it in the c code.

Guo, Yejun yejun.guo at intel.com
Mon Jul 29 04:56:54 EEST 2019


since tf.pad is enabled, the conv2d(valid) changes back to its original behavior.

Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
---
 libavfilter/dnn/dnn_backend_native.c    | 35 +++++++++++++++++++++++++++++++++
 libavfilter/dnn/dnn_backend_native.h    |  2 +-
 tools/python/convert_from_tensorflow.py | 23 +++++++++++++++++-----
 3 files changed, 54 insertions(+), 6 deletions(-)

diff --git a/libavfilter/dnn/dnn_backend_native.c b/libavfilter/dnn/dnn_backend_native.c
index 82e900b..09c583b 100644
--- a/libavfilter/dnn/dnn_backend_native.c
+++ b/libavfilter/dnn/dnn_backend_native.c
@@ -25,6 +25,7 @@
 
 #include "dnn_backend_native.h"
 #include "libavutil/avassert.h"
+#include "dnn_backend_native_layer_pad.h"
 
 static DNNReturnType set_input_output_native(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output)
 {
@@ -32,6 +33,7 @@ static DNNReturnType set_input_output_native(void *model, DNNInputData *input, c
     InputParams *input_params;
     ConvolutionalParams *conv_params;
     DepthToSpaceParams *depth_to_space_params;
+    LayerPadParams *pad_params;
     int cur_width, cur_height, cur_channels;
     int32_t layer;
 
@@ -77,6 +79,12 @@ static DNNReturnType set_input_output_native(void *model, DNNInputData *input, c
             cur_height *= depth_to_space_params->block_size;
             cur_width *= depth_to_space_params->block_size;
             break;
+        case MIRROR_PAD:
+            pad_params = (LayerPadParams *)network->layers[layer].params;
+            cur_height = cur_height + pad_params->paddings[1][0] + pad_params->paddings[1][1];
+            cur_width = cur_width + pad_params->paddings[2][0] + pad_params->paddings[2][1];
+            cur_channels = cur_channels + pad_params->paddings[3][0] + pad_params->paddings[3][1];
+            break;
         default:
             return DNN_ERROR;
         }
@@ -110,6 +118,7 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename)
     DNNLayerType layer_type;
     ConvolutionalParams *conv_params;
     DepthToSpaceParams *depth_to_space_params;
+    LayerPadParams *pad_params;
 
     model = av_malloc(sizeof(DNNModel));
     if (!model){
@@ -207,6 +216,23 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename)
             network->layers[layer].type = DEPTH_TO_SPACE;
             network->layers[layer].params = depth_to_space_params;
             break;
+        case MIRROR_PAD:
+            pad_params = av_malloc(sizeof(LayerPadParams));
+            if (!pad_params){
+                avio_closep(&model_file_context);
+                ff_dnn_free_model_native(&model);
+                return NULL;
+            }
+            pad_params->mode = (int32_t)avio_rl32(model_file_context);
+            dnn_size += 4;
+            for (i = 0; i < 4; ++i) {
+                pad_params->paddings[i][0] = avio_rl32(model_file_context);
+                pad_params->paddings[i][1] = avio_rl32(model_file_context);
+                dnn_size += 8;
+            }
+            network->layers[layer].type = MIRROR_PAD;
+            network->layers[layer].params = pad_params;
+            break;
         default:
             avio_closep(&model_file_context);
             ff_dnn_free_model_native(&model);
@@ -314,6 +340,7 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
     InputParams *input_params;
     ConvolutionalParams *conv_params;
     DepthToSpaceParams *depth_to_space_params;
+    LayerPadParams *pad_params;
 
     if (network->layers_num <= 0 || network->layers[0].type != INPUT || !network->layers[0].output){
         return DNN_ERROR;
@@ -348,6 +375,14 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
             cur_width *= depth_to_space_params->block_size;
             cur_channels /= depth_to_space_params->block_size * depth_to_space_params->block_size;
             break;
+        case MIRROR_PAD:
+            pad_params = (LayerPadParams *)network->layers[layer].params;
+            dnn_execute_layer_pad(network->layers[layer - 1].output, network->layers[layer].output,
+                                  pad_params, 1, cur_height, cur_width, cur_channels);
+            cur_height = cur_height + pad_params->paddings[1][0] + pad_params->paddings[1][1];
+            cur_width = cur_width + pad_params->paddings[2][0] + pad_params->paddings[2][1];
+            cur_channels = cur_channels + pad_params->paddings[3][0] + pad_params->paddings[3][1];
+            break;
         case INPUT:
             return DNN_ERROR;
         }
diff --git a/libavfilter/dnn/dnn_backend_native.h b/libavfilter/dnn/dnn_backend_native.h
index 8ef1855..b6f9533 100644
--- a/libavfilter/dnn/dnn_backend_native.h
+++ b/libavfilter/dnn/dnn_backend_native.h
@@ -30,7 +30,7 @@
 #include "../dnn_interface.h"
 #include "libavformat/avio.h"
 
-typedef enum {INPUT, CONV, DEPTH_TO_SPACE} DNNLayerType;
+typedef enum {INPUT, CONV, DEPTH_TO_SPACE, MIRROR_PAD} DNNLayerType;
 
 typedef enum {RELU, TANH, SIGMOID, NONE, LEAKY_RELU} DNNActivationFunc;
 
diff --git a/tools/python/convert_from_tensorflow.py b/tools/python/convert_from_tensorflow.py
index 37049e5..041c82c 100644
--- a/tools/python/convert_from_tensorflow.py
+++ b/tools/python/convert_from_tensorflow.py
@@ -23,9 +23,6 @@ import sys, struct
 
 __all__ = ['convert_from_tensorflow']
 
-# as the first step to be compatible with vf_sr, it is not general.
-# it will be refined step by step.
-
 class TFConverter:
     def __init__(self, graph_def, nodes, outfile):
         self.graph_def = graph_def
@@ -36,9 +33,10 @@ class TFConverter:
         self.name_node_dict = {}
         self.edges = {}
         self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2, 'LeakyRelu':4}
-        self.conv_paddings = {'VALID':2, 'SAME':1}
+        self.conv_paddings = {'VALID':0, 'SAME':1}
         self.converted_nodes = set()
-        self.op2code = {'Conv2D':1, 'DepthToSpace':2}
+        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3}
+        self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
 
 
     def dump_for_tensorboard(self):
@@ -101,6 +99,19 @@ class TFConverter:
         self.converted_nodes.add(node.name)
 
 
+    def dump_mirrorpad_to_file(self, node, f):
+        assert(node.op == 'MirrorPad')
+        self.layer_number = self.layer_number + 1
+        mode = node.attr['mode'].s
+        mode = self.mirrorpad_mode[mode.decode("utf-8")]
+        np.array([self.op2code[node.op], mode], dtype=np.uint32).tofile(f)
+        pnode = self.name_node_dict[node.input[1]]
+        self.converted_nodes.add(pnode.name)
+        paddings = pnode.attr['value'].tensor.tensor_content
+        f.write(paddings)
+        self.converted_nodes.add(node.name)
+
+
     def generate_layer_number(self):
         # in current hard code implementation, the layer number is the first data written to the native model file
         # it is not easy to know it at the beginning time in the general converter, so first do a dry run for compatibility
@@ -118,6 +129,8 @@ class TFConverter:
                 self.dump_conv2d_to_file(node, f)
             elif node.op == 'DepthToSpace':
                 self.dump_depth2space_to_file(node, f)
+            elif node.op == 'MirrorPad':
+                self.dump_mirrorpad_to_file(node, f)
 
 
     def dump_to_file(self):
-- 
2.7.4



More information about the ffmpeg-devel mailing list